COGNITIVE_MODEL.md

MemoryKit Cognitive Model

Overview

MemoryKit models human cognition to create efficient memory systems for LLM applications. This document explains the neuroscience-inspired architecture.

Brain Component Mapping

1. Prefrontal Cortex → PrefrontalController

Role: Executive function and attention control

The prefrontal cortex:

• Makes decisions about what to focus on
• Plans complex behaviors
• Evaluates alternatives
• Manages working memory

Software Implementation:

public interface IPrefrontalController
{
    Task<QueryPlan> BuildQueryPlanAsync(string query, ConversationState state);
    Task<QueryType> ClassifyQueryAsync(string query);
    List<MemoryLayer> DetermineLayersToUse(QueryType type, ConversationState state);
}

Query Planning Strategy:

• Continuation → Layer 3 only (fast)
• Fact Retrieval → Layers 2-3 (balanced)
• Deep Recall → Layers 1-3 (thorough)
• Complex → All layers (comprehensive)
• Procedural → Layers 3 + P (routine)

2. Amygdala → AmygdalaImportanceEngine

Role: Emotional tagging and importance scoring

The amygdala:

• Tags experiences with emotional significance
• Influences what gets consolidated to long-term memory
• Modulates memory strength based on emotional arousal

Software Implementation:

public interface IAmygdalaImportanceEngine : IImportanceEngine
{
    Task<double> CalculateImportanceAsync(Message message);
    Task<(double Score, string Sentiment)> AnalyzeSentimentAsync(string text);
    bool ContainsDecisionLanguage(string text);
    bool HasExplicitImportanceMarkers(string text);
}

Importance Scoring Algorithm:

FinalScore = (BaseScore × 0.4) + 
             (EmotionalWeight × 0.3) + 
             (NoveltyBoost × 0.2) + 
             (RecencyFactor × 0.1)

Importance Triggers:

• User questions (1.5x boost)
• Decisions/commitments (2.0x boost)
• Emotional language (1.3x boost)
• Novel information (1.5x boost)
• Explicit importance markers (2.5x boost)

3. Hippocampus → HippocampusIndexer

Role: Temporary storage and consolidation initiation

The hippocampus:

• Rapidly encodes new information
• Indexes memories for later retrieval
• Initiates consolidation to cortical storage

Software Implementation:

public interface IHippocampusIndexer
{
    Task<string> EncodeAsync(Message message);
    Task MarkForConsolidationAsync(string messageId);
    Task ConsolidateAsync(string userId);
}

Consolidation Process:

1. Encoding: Initial capture in working memory
2. Indexing: Create search indices
3. Importance Scoring: Amygdala processing
4. Consolidation: Move to appropriate layer
   • High importance → Layer 1 (episodic)
   • Facts → Layer 2 (semantic)
   • Patterns → Layer P (procedural)

4. Basal Ganglia → ProceduralMemoryService

Role: Procedural memory and habitual responses

The basal ganglia:

• Store learned procedures and routines
• Execute habitual behaviors
• Learn from reward and feedback

Software Implementation:

public interface IProceduralMemoryService
{
    Task<ProceduralPattern?> MatchPatternAsync(string userId, string query);
    Task DetectAndStorePatternAsync(string userId, Message message);
    Task<ProceduralPattern[]> GetUserPatternsAsync(string userId);
}

Pattern Learning:

• Detect: Extract rules from user instructions
• Store: Save as procedural patterns
• Reinforce: Increase confidence with each use
• Decay: Reduce confidence if unused

5. Neocortex → ScratchpadService

Role: Long-term semantic knowledge

The neocortex:

• Stores consolidated semantic knowledge
• Integrates information across domains
• Supports reasoning and generalization

Software Implementation: Azure Table Storage with semantic indexing

Memory System

Working Memory (Layer 3)

Capacity: ~7±2 items Substrate: Redis cache Latency: <5ms Duration: ~30 seconds to minutes Function: Active processing

Semantic Memory (Layer 2)

Capacity: Unlimited Substrate: Azure Table Storage + embeddings Latency: ~30ms Duration: Long-term Function: Facts, concepts, relationships

Episodic Memory (Layer 1)

Capacity: Full history Substrate: Azure Blob + AI Search Latency: ~120ms Duration: Long-term Function: Specific events and experiences

Procedural Memory (Layer P)

Capacity: Learned patterns Substrate: Azure Table Storage Latency: ~50ms Duration: Long-term Function: Skills and routines

Sleep-Based Consolidation

Unlike human sleep, MemoryKit performs continuous consolidation:

Event → Working Memory → Importance Scoring → 
Consolidation Decision → Target Layer Storage

Consolidation Rules:

• High importance + specific facts → Layer 1
• General knowledge → Layer 2
• Repeated procedures → Layer P
• Low importance → Discard after TTL

Attention and Filtering

Selective Attention: The PrefrontalController manages attention by:

1. Classifying query type
2. Determining relevant layers
3. Filtering irrelevant information
4. Assembling minimal context

Token Efficiency:

• Continuation: ~100 tokens
• Fact Retrieval: ~300-500 tokens
• Deep Recall: ~1000-1500 tokens
• Complex: ~2000 tokens

Learning and Adaptation

Reinforcement Learning

• Procedural patterns increase confidence with use
• Importance thresholds adjust based on retention
• Layer selection optimizes based on query latency

Semantic Learning

• Entity embeddings capture relationships
• Fact importance updates as accessed
• Procedural rules strengthen with application

Emotional Dimension

Emotional Arousal: Increases importance

• Explicit markers: "important", "critical"
• Sentiment analysis: high positive/negative
• Decision language: commitments, promises
• Novelty: new information

Emotional Decay: Importance fades over time

• Recent events more salient
• Exponential decay function
• Can be reinforced through repeated access

Psychological Principles

1. Spacing Effect

• Repeated access reinforces memories
• Intervals optimize consolidation
• Implemented through access tracking

2. Recency Effect

• Recent items prioritized
• Reflected in working memory LRU
• Temporal weighting in retrieval

3. Primacy Effect

• First occurrences marked as novel
• Initial importance boost
• Entity tracking captures first mention

4. Reconstructive Memory

• Context influences recall
• MemoryContext reassembles information
• Query influences what's retrieved

5. Transfer of Learning

• Procedural patterns enable transfer
• Semantic knowledge generalizes
• Relationships captured in embeddings

Disorders & Failure Modes

Working Memory Overflow

Problem: Too many recent items Solution: LRU eviction, importance-based retention

Semantic Degradation

Problem: False or outdated facts Solution: Versioning, update tracking, confidence scores

Procedural Rigidity

Problem: Patterns too strict, not adapting Solution: Dynamic confidence thresholds, decay for unused patterns

Consolidation Failure

Problem: Important information lost Solution: Redundant storage, importance-based TTL

Future Enhancements

1. Metacognition

• Monitor own performance
• Adjust learning strategies
• Self-aware uncertainty

2. Emotional Regulation

• Learn from emotional responses
• Adjust importance scoring over time
• Empathetic responses

3. Executive Function Improvement

• Better query planning
• Confidence calibration
• Multi-step reasoning

4. Transfer Learning

• Learn across domains
• Apply patterns broadly
• Generalize from experience

References

Neuroscience

• Kandel, E.R., et al. "Principles of Neural Science"
• LeDoux, J. "The Emotional Brain"
• Squire, L.R. "Memory and the Hippocampus"

Cognitive Psychology

• Baddeley, A. "Working Memory: Theories, Models, and Controversies"
• Tulving, E. "Elements of Episodic Memory"
• Anderson, J.R. "Cognitive Psychology and its Implications"

AI/LLM Applications

• Vaswani, A., et al. "Attention is All You Need"
• Brown, T.B., et al. "Language Models are Few-Shot Learners"
• Wei, J., et al. "Emergent Abilities of Large Language Models"